git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_iter.h

   1 /* SPDX-License-Identifier: GPL-2.0 */
   2 #ifndef _BCACHEFS_BTREE_ITER_H
   3 #define _BCACHEFS_BTREE_ITER_H
   4
   5 #include "bset.h"
   6 #include "btree_types.h"
   7 #include "trace.h"
   8
   9 static inline int __bkey_err(const struct bkey *k)
  10 {
  11         return PTR_ERR_OR_ZERO(k);
  12 }
  13
  14 #define bkey_err(_k)    __bkey_err((_k).k)
  15
  16 static inline void __btree_path_get(struct btree_path *path, bool intent)
  17 {
  18         path->ref++;
  19         path->intent_ref += intent;
  20 }
  21
  22 static inline bool __btree_path_put(struct btree_path *path, bool intent)
  23 {
  24         EBUG_ON(!path->ref);
  25         EBUG_ON(!path->intent_ref && intent);
  26         path->intent_ref -= intent;
  27         return --path->ref == 0;
  28 }
  29
  30 static inline void btree_path_set_dirty(struct btree_path *path,
  31                                         enum btree_path_uptodate u)
  32 {
  33         path->uptodate = max_t(unsigned, path->uptodate, u);
  34 }
  35
  36 static inline struct btree *btree_path_node(struct btree_path *path,
  37                                             unsigned level)
  38 {
  39         return level < BTREE_MAX_DEPTH ? path->l[level].b : NULL;
  40 }
  41
  42 static inline bool btree_node_lock_seq_matches(const struct btree_path *path,
  43                                         const struct btree *b, unsigned level)
  44 {
  45         return path->l[level].lock_seq == six_lock_seq(&b->c.lock);
  46 }
  47
  48 static inline struct btree *btree_node_parent(struct btree_path *path,
  49                                               struct btree *b)
  50 {
  51         return btree_path_node(path, b->c.level + 1);
  52 }
  53
  54 /* Iterate over paths within a transaction: */
  55
  56 void __bch2_btree_trans_sort_paths(struct btree_trans *);
  57
  58 static inline void btree_trans_sort_paths(struct btree_trans *trans)
  59 {
  60         if (!IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
  61             trans->paths_sorted)
  62                 return;
  63         __bch2_btree_trans_sort_paths(trans);
  64 }
  65
  66 static inline struct btree_path *
  67 __trans_next_path(struct btree_trans *trans, unsigned idx)
  68 {
  69         u64 l;
  70
  71         if (idx == BTREE_ITER_MAX)
  72                 return NULL;
  73
  74         l = trans->paths_allocated >> idx;
  75         if (!l)
  76                 return NULL;
  77
  78         idx += __ffs64(l);
  79         EBUG_ON(idx >= BTREE_ITER_MAX);
  80         EBUG_ON(trans->paths[idx].idx != idx);
  81         return &trans->paths[idx];
  82 }
  83
  84 #define trans_for_each_path_from(_trans, _path, _start)                 \
  85         for (_path = __trans_next_path((_trans), _start);               \
  86              (_path);                                                   \
  87              _path = __trans_next_path((_trans), (_path)->idx + 1))
  88
  89 #define trans_for_each_path(_trans, _path)                              \
  90         trans_for_each_path_from(_trans, _path, 0)
  91
  92 static inline struct btree_path *
  93 __trans_next_path_safe(struct btree_trans *trans, unsigned *idx)
  94 {
  95         u64 l;
  96
  97         if (*idx == BTREE_ITER_MAX)
  98                 return NULL;
  99
 100         l = trans->paths_allocated >> *idx;
 101         if (!l)
 102                 return NULL;
 103
 104         *idx += __ffs64(l);
 105         EBUG_ON(*idx >= BTREE_ITER_MAX);
 106         return &trans->paths[*idx];
 107 }
 108
 109 /*
 110  * This version is intended to be safe for use on a btree_trans that is owned by
 111  * another thread, for bch2_btree_trans_to_text();
 112  */
 113 #define trans_for_each_path_safe_from(_trans, _path, _idx, _start)      \
 114         for (_idx = _start;                                             \
 115              (_path = __trans_next_path_safe((_trans), &_idx));         \
 116              _idx++)
 117
 118 #define trans_for_each_path_safe(_trans, _path, _idx)                   \
 119         trans_for_each_path_safe_from(_trans, _path, _idx, 0)
 120
 121 static inline struct btree_path *next_btree_path(struct btree_trans *trans, struct btree_path *path)
 122 {
 123         unsigned idx = path ? path->sorted_idx + 1 : 0;
 124
 125         EBUG_ON(idx > trans->nr_sorted);
 126
 127         return idx < trans->nr_sorted
 128                 ? trans->paths + trans->sorted[idx]
 129                 : NULL;
 130 }
 131
 132 static inline struct btree_path *prev_btree_path(struct btree_trans *trans, struct btree_path *path)
 133 {
 134         unsigned idx = path ? path->sorted_idx : trans->nr_sorted;
 135
 136         return idx
 137                 ? trans->paths + trans->sorted[idx - 1]
 138                 : NULL;
 139 }
 140
 141 #define trans_for_each_path_inorder(_trans, _path, _i)                  \
 142         for (_i = 0;                                                    \
 143              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) < (_trans)->nr_sorted;\
 144              _i++)
 145
 146 #define trans_for_each_path_inorder_reverse(_trans, _path, _i)          \
 147         for (_i = trans->nr_sorted - 1;                                 \
 148              ((_path) = (_trans)->paths + trans->sorted[_i]), (_i) >= 0;\
 149              --_i)
 150
 151 static inline bool __path_has_node(const struct btree_path *path,
 152                                    const struct btree *b)
 153 {
 154         return path->l[b->c.level].b == b &&
 155                 btree_node_lock_seq_matches(path, b, b->c.level);
 156 }
 157
 158 static inline struct btree_path *
 159 __trans_next_path_with_node(struct btree_trans *trans, struct btree *b,
 160                             unsigned idx)
 161 {
 162         struct btree_path *path = __trans_next_path(trans, idx);
 163
 164         while (path && !__path_has_node(path, b))
 165                 path = __trans_next_path(trans, path->idx + 1);
 166
 167         return path;
 168 }
 169
 170 #define trans_for_each_path_with_node(_trans, _b, _path)                \
 171         for (_path = __trans_next_path_with_node((_trans), (_b), 0);    \
 172              (_path);                                                   \
 173              _path = __trans_next_path_with_node((_trans), (_b),        \
 174                                                  (_path)->idx + 1))
 175
 176 struct btree_path *__bch2_btree_path_make_mut(struct btree_trans *, struct btree_path *,
 177                          bool, unsigned long);
 178
 179 static inline struct btree_path * __must_check
 180 bch2_btree_path_make_mut(struct btree_trans *trans,
 181                          struct btree_path *path, bool intent,
 182                          unsigned long ip)
 183 {
 184         if (path->ref > 1 || path->preserve)
 185                 path = __bch2_btree_path_make_mut(trans, path, intent, ip);
 186         path->should_be_locked = false;
 187         return path;
 188 }
 189
 190 struct btree_path * __must_check
 191 __bch2_btree_path_set_pos(struct btree_trans *, struct btree_path *,
 192                         struct bpos, bool, unsigned long, int);
 193
 194 static inline struct btree_path * __must_check
 195 bch2_btree_path_set_pos(struct btree_trans *trans,
 196                    struct btree_path *path, struct bpos new_pos,
 197                    bool intent, unsigned long ip)
 198 {
 199         int cmp = bpos_cmp(new_pos, path->pos);
 200
 201         return cmp
 202                 ? __bch2_btree_path_set_pos(trans, path, new_pos, intent, ip, cmp)
 203                 : path;
 204 }
 205
 206 int __must_check bch2_btree_path_traverse_one(struct btree_trans *, struct btree_path *,
 207                                               unsigned, unsigned long);
 208
 209 static inline int __must_check bch2_btree_path_traverse(struct btree_trans *trans,
 210                                           struct btree_path *path, unsigned flags)
 211 {
 212         if (path->uptodate < BTREE_ITER_NEED_RELOCK)
 213                 return 0;
 214
 215         return bch2_btree_path_traverse_one(trans, path, flags, _RET_IP_);
 216 }
 217
 218 int __must_check bch2_btree_path_traverse(struct btree_trans *,
 219                                           struct btree_path *, unsigned);
 220 struct btree_path *bch2_path_get(struct btree_trans *, enum btree_id, struct bpos,
 221                                  unsigned, unsigned, unsigned, unsigned long);
 222 struct bkey_s_c bch2_btree_path_peek_slot(struct btree_path *, struct bkey *);
 223
 224 /*
 225  * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
 226  * different snapshot:
 227  */
 228 static inline struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
 229 {
 230         struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
 231
 232         if (k.k && bpos_eq(path->pos, k.k->p))
 233                 return k;
 234
 235         bkey_init(u);
 236         u->p = path->pos;
 237         return (struct bkey_s_c) { u, NULL };
 238 }
 239
 240 struct bkey_i *bch2_btree_journal_peek_slot(struct btree_trans *,
 241                                         struct btree_iter *, struct bpos);
 242
 243 void bch2_btree_path_level_init(struct btree_trans *, struct btree_path *, struct btree *);
 244
 245 int __bch2_trans_mutex_lock(struct btree_trans *, struct mutex *);
 246
 247 static inline int bch2_trans_mutex_lock(struct btree_trans *trans, struct mutex *lock)
 248 {
 249         return mutex_trylock(lock)
 250                 ? 0
 251                 : __bch2_trans_mutex_lock(trans, lock);
 252 }
 253
 254 #ifdef CONFIG_BCACHEFS_DEBUG
 255 void bch2_trans_verify_paths(struct btree_trans *);
 256 void bch2_assert_pos_locked(struct btree_trans *, enum btree_id,
 257                             struct bpos, bool);
 258 #else
 259 static inline void bch2_trans_verify_paths(struct btree_trans *trans) {}
 260 static inline void bch2_assert_pos_locked(struct btree_trans *trans, enum btree_id id,
 261                                           struct bpos pos, bool key_cache) {}
 262 #endif
 263
 264 void bch2_btree_path_fix_key_modified(struct btree_trans *trans,
 265                                       struct btree *, struct bkey_packed *);
 266 void bch2_btree_node_iter_fix(struct btree_trans *trans, struct btree_path *,
 267                               struct btree *, struct btree_node_iter *,
 268                               struct bkey_packed *, unsigned, unsigned);
 269
 270 int bch2_btree_path_relock_intent(struct btree_trans *, struct btree_path *);
 271
 272 void bch2_path_put(struct btree_trans *, struct btree_path *, bool);
 273
 274 int bch2_trans_relock(struct btree_trans *);
 275 int bch2_trans_relock_notrace(struct btree_trans *);
 276 void bch2_trans_unlock(struct btree_trans *);
 277 void bch2_trans_unlock_long(struct btree_trans *);
 278 bool bch2_trans_locked(struct btree_trans *);
 279
 280 static inline int trans_was_restarted(struct btree_trans *trans, u32 restart_count)
 281 {
 282         return restart_count != trans->restart_count
 283                 ? -BCH_ERR_transaction_restart_nested
 284                 : 0;
 285 }
 286
 287 void __noreturn bch2_trans_restart_error(struct btree_trans *, u32);
 288
 289 static inline void bch2_trans_verify_not_restarted(struct btree_trans *trans,
 290                                                    u32 restart_count)
 291 {
 292         if (trans_was_restarted(trans, restart_count))
 293                 bch2_trans_restart_error(trans, restart_count);
 294 }
 295
 296 void __noreturn bch2_trans_in_restart_error(struct btree_trans *);
 297
 298 static inline void bch2_trans_verify_not_in_restart(struct btree_trans *trans)
 299 {
 300         if (trans->restarted)
 301                 bch2_trans_in_restart_error(trans);
 302 }
 303
 304 __always_inline
 305 static int btree_trans_restart_nounlock(struct btree_trans *trans, int err)
 306 {
 307         BUG_ON(err <= 0);
 308         BUG_ON(!bch2_err_matches(-err, BCH_ERR_transaction_restart));
 309
 310         trans->restarted = err;
 311         trans->last_restarted_ip = _THIS_IP_;
 312         return -err;
 313 }
 314
 315 __always_inline
 316 static int btree_trans_restart(struct btree_trans *trans, int err)
 317 {
 318         btree_trans_restart_nounlock(trans, err);
 319         return -err;
 320 }
 321
 322 bool bch2_btree_node_upgrade(struct btree_trans *,
 323                              struct btree_path *, unsigned);
 324
 325 void __bch2_btree_path_downgrade(struct btree_trans *, struct btree_path *, unsigned);
 326
 327 static inline void bch2_btree_path_downgrade(struct btree_trans *trans,
 328                                              struct btree_path *path)
 329 {
 330         unsigned new_locks_want = path->level + !!path->intent_ref;
 331
 332         if (path->locks_want > new_locks_want)
 333                 __bch2_btree_path_downgrade(trans, path, new_locks_want);
 334 }
 335
 336 void bch2_trans_downgrade(struct btree_trans *);
 337
 338 void bch2_trans_node_add(struct btree_trans *trans, struct btree *);
 339 void bch2_trans_node_reinit_iter(struct btree_trans *, struct btree *);
 340
 341 int __must_check __bch2_btree_iter_traverse(struct btree_iter *iter);
 342 int __must_check bch2_btree_iter_traverse(struct btree_iter *);
 343
 344 struct btree *bch2_btree_iter_peek_node(struct btree_iter *);
 345 struct btree *bch2_btree_iter_peek_node_and_restart(struct btree_iter *);
 346 struct btree *bch2_btree_iter_next_node(struct btree_iter *);
 347
 348 struct bkey_s_c bch2_btree_iter_peek_upto(struct btree_iter *, struct bpos);
 349 struct bkey_s_c bch2_btree_iter_next(struct btree_iter *);
 350
 351 static inline struct bkey_s_c bch2_btree_iter_peek(struct btree_iter *iter)
 352 {
 353         return bch2_btree_iter_peek_upto(iter, SPOS_MAX);
 354 }
 355
 356 struct bkey_s_c bch2_btree_iter_peek_prev(struct btree_iter *);
 357 struct bkey_s_c bch2_btree_iter_prev(struct btree_iter *);
 358
 359 struct bkey_s_c bch2_btree_iter_peek_slot(struct btree_iter *);
 360 struct bkey_s_c bch2_btree_iter_next_slot(struct btree_iter *);
 361 struct bkey_s_c bch2_btree_iter_prev_slot(struct btree_iter *);
 362
 363 bool bch2_btree_iter_advance(struct btree_iter *);
 364 bool bch2_btree_iter_rewind(struct btree_iter *);
 365
 366 static inline void __bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 367 {
 368         iter->k.type = KEY_TYPE_deleted;
 369         iter->k.p.inode         = iter->pos.inode       = new_pos.inode;
 370         iter->k.p.offset        = iter->pos.offset      = new_pos.offset;
 371         iter->k.p.snapshot      = iter->pos.snapshot    = new_pos.snapshot;
 372         iter->k.size = 0;
 373 }
 374
 375 static inline void bch2_btree_iter_set_pos(struct btree_iter *iter, struct bpos new_pos)
 376 {
 377         if (unlikely(iter->update_path))
 378                 bch2_path_put(iter->trans, iter->update_path,
 379                               iter->flags & BTREE_ITER_INTENT);
 380         iter->update_path = NULL;
 381
 382         if (!(iter->flags & BTREE_ITER_ALL_SNAPSHOTS))
 383                 new_pos.snapshot = iter->snapshot;
 384
 385         __bch2_btree_iter_set_pos(iter, new_pos);
 386 }
 387
 388 static inline void bch2_btree_iter_set_pos_to_extent_start(struct btree_iter *iter)
 389 {
 390         BUG_ON(!(iter->flags & BTREE_ITER_IS_EXTENTS));
 391         iter->pos = bkey_start_pos(&iter->k);
 392 }
 393
 394 static inline void bch2_btree_iter_set_snapshot(struct btree_iter *iter, u32 snapshot)
 395 {
 396         struct bpos pos = iter->pos;
 397
 398         iter->snapshot = snapshot;
 399         pos.snapshot = snapshot;
 400         bch2_btree_iter_set_pos(iter, pos);
 401 }
 402
 403 void bch2_trans_iter_exit(struct btree_trans *, struct btree_iter *);
 404
 405 static inline unsigned __bch2_btree_iter_flags(struct btree_trans *trans,
 406                                                unsigned btree_id,
 407                                                unsigned flags)
 408 {
 409         if (!(flags & (BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_NOT_EXTENTS)) &&
 410             btree_id_is_extents(btree_id))
 411                 flags |= BTREE_ITER_IS_EXTENTS;
 412
 413         if (!(flags & __BTREE_ITER_ALL_SNAPSHOTS) &&
 414             !btree_type_has_snapshot_field(btree_id))
 415                 flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
 416
 417         if (!(flags & BTREE_ITER_ALL_SNAPSHOTS) &&
 418             btree_type_has_snapshots(btree_id))
 419                 flags |= BTREE_ITER_FILTER_SNAPSHOTS;
 420
 421         if (trans->journal_replay_not_finished)
 422                 flags |= BTREE_ITER_WITH_JOURNAL;
 423
 424         return flags;
 425 }
 426
 427 static inline unsigned bch2_btree_iter_flags(struct btree_trans *trans,
 428                                              unsigned btree_id,
 429                                              unsigned flags)
 430 {
 431         if (!btree_id_cached(trans->c, btree_id)) {
 432                 flags &= ~BTREE_ITER_CACHED;
 433                 flags &= ~BTREE_ITER_WITH_KEY_CACHE;
 434         } else if (!(flags & BTREE_ITER_CACHED))
 435                 flags |= BTREE_ITER_WITH_KEY_CACHE;
 436
 437         return __bch2_btree_iter_flags(trans, btree_id, flags);
 438 }
 439
 440 static inline void bch2_trans_iter_init_common(struct btree_trans *trans,
 441                                           struct btree_iter *iter,
 442                                           unsigned btree_id, struct bpos pos,
 443                                           unsigned locks_want,
 444                                           unsigned depth,
 445                                           unsigned flags,
 446                                           unsigned long ip)
 447 {
 448         iter->trans             = trans;
 449         iter->update_path       = NULL;
 450         iter->key_cache_path    = NULL;
 451         iter->btree_id          = btree_id;
 452         iter->min_depth         = 0;
 453         iter->flags             = flags;
 454         iter->snapshot          = pos.snapshot;
 455         iter->pos               = pos;
 456         iter->k                 = POS_KEY(pos);
 457         iter->journal_idx       = 0;
 458 #ifdef CONFIG_BCACHEFS_DEBUG
 459         iter->ip_allocated = ip;
 460 #endif
 461         iter->path = bch2_path_get(trans, btree_id, iter->pos,
 462                                    locks_want, depth, flags, ip);
 463 }
 464
 465 void bch2_trans_iter_init_outlined(struct btree_trans *, struct btree_iter *,
 466                           enum btree_id, struct bpos, unsigned);
 467
 468 static inline void bch2_trans_iter_init(struct btree_trans *trans,
 469                           struct btree_iter *iter,
 470                           unsigned btree_id, struct bpos pos,
 471                           unsigned flags)
 472 {
 473         if (__builtin_constant_p(btree_id) &&
 474             __builtin_constant_p(flags))
 475                 bch2_trans_iter_init_common(trans, iter, btree_id, pos, 0, 0,
 476                                 bch2_btree_iter_flags(trans, btree_id, flags),
 477                                 _THIS_IP_);
 478         else
 479                 bch2_trans_iter_init_outlined(trans, iter, btree_id, pos, flags);
 480 }
 481
 482 void bch2_trans_node_iter_init(struct btree_trans *, struct btree_iter *,
 483                                enum btree_id, struct bpos,
 484                                unsigned, unsigned, unsigned);
 485 void bch2_trans_copy_iter(struct btree_iter *, struct btree_iter *);
 486
 487 static inline void set_btree_iter_dontneed(struct btree_iter *iter)
 488 {
 489         if (!iter->trans->restarted)
 490                 iter->path->preserve = false;
 491 }
 492
 493 void *__bch2_trans_kmalloc(struct btree_trans *, size_t);
 494
 495 static inline void *bch2_trans_kmalloc(struct btree_trans *trans, size_t size)
 496 {
 497         size = roundup(size, 8);
 498
 499         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 500                 void *p = trans->mem + trans->mem_top;
 501
 502                 trans->mem_top += size;
 503                 memset(p, 0, size);
 504                 return p;
 505         } else {
 506                 return __bch2_trans_kmalloc(trans, size);
 507         }
 508 }
 509
 510 static inline void *bch2_trans_kmalloc_nomemzero(struct btree_trans *trans, size_t size)
 511 {
 512         size = roundup(size, 8);
 513
 514         if (likely(trans->mem_top + size <= trans->mem_bytes)) {
 515                 void *p = trans->mem + trans->mem_top;
 516
 517                 trans->mem_top += size;
 518                 return p;
 519         } else {
 520                 return __bch2_trans_kmalloc(trans, size);
 521         }
 522 }
 523
 524 static inline struct bkey_s_c __bch2_bkey_get_iter(struct btree_trans *trans,
 525                                 struct btree_iter *iter,
 526                                 unsigned btree_id, struct bpos pos,
 527                                 unsigned flags, unsigned type)
 528 {
 529         struct bkey_s_c k;
 530
 531         bch2_trans_iter_init(trans, iter, btree_id, pos, flags);
 532         k = bch2_btree_iter_peek_slot(iter);
 533
 534         if (!bkey_err(k) && type && k.k->type != type)
 535                 k = bkey_s_c_err(-BCH_ERR_ENOENT_bkey_type_mismatch);
 536         if (unlikely(bkey_err(k)))
 537                 bch2_trans_iter_exit(trans, iter);
 538         return k;
 539 }
 540
 541 static inline struct bkey_s_c bch2_bkey_get_iter(struct btree_trans *trans,
 542                                 struct btree_iter *iter,
 543                                 unsigned btree_id, struct bpos pos,
 544                                 unsigned flags)
 545 {
 546         return __bch2_bkey_get_iter(trans, iter, btree_id, pos, flags, 0);
 547 }
 548
 549 #define bch2_bkey_get_iter_typed(_trans, _iter, _btree_id, _pos, _flags, _type)\
 550         bkey_s_c_to_##_type(__bch2_bkey_get_iter(_trans, _iter,                 \
 551                                        _btree_id, _pos, _flags, KEY_TYPE_##_type))
 552
 553 static inline int __bch2_bkey_get_val_typed(struct btree_trans *trans,
 554                                 unsigned btree_id, struct bpos pos,
 555                                 unsigned flags, unsigned type,
 556                                 unsigned val_size, void *val)
 557 {
 558         struct btree_iter iter;
 559         struct bkey_s_c k;
 560         int ret;
 561
 562         k = __bch2_bkey_get_iter(trans, &iter, btree_id, pos, flags, type);
 563         ret = bkey_err(k);
 564         if (!ret) {
 565                 unsigned b = min_t(unsigned, bkey_val_bytes(k.k), val_size);
 566
 567                 memcpy(val, k.v, b);
 568                 if (unlikely(b < sizeof(*val)))
 569                         memset((void *) val + b, 0, sizeof(*val) - b);
 570                 bch2_trans_iter_exit(trans, &iter);
 571         }
 572
 573         return ret;
 574 }
 575
 576 #define bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags, _type, _val)\
 577         __bch2_bkey_get_val_typed(_trans, _btree_id, _pos, _flags,      \
 578                                   KEY_TYPE_##_type, sizeof(*_val), _val)
 579
 580 void bch2_trans_srcu_unlock(struct btree_trans *);
 581 void bch2_trans_srcu_lock(struct btree_trans *);
 582
 583 u32 bch2_trans_begin(struct btree_trans *);
 584
 585 /*
 586  * XXX
 587  * this does not handle transaction restarts from bch2_btree_iter_next_node()
 588  * correctly
 589  */
 590 #define __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 591                               _locks_want, _depth, _flags, _b, _ret)    \
 592         for (bch2_trans_node_iter_init((_trans), &(_iter), (_btree_id), \
 593                                 _start, _locks_want, _depth, _flags);   \
 594              (_b) = bch2_btree_iter_peek_node_and_restart(&(_iter)),    \
 595              !((_ret) = PTR_ERR_OR_ZERO(_b)) && (_b);                   \
 596              (_b) = bch2_btree_iter_next_node(&(_iter)))
 597
 598 #define for_each_btree_node(_trans, _iter, _btree_id, _start,           \
 599                             _flags, _b, _ret)                           \
 600         __for_each_btree_node(_trans, _iter, _btree_id, _start,         \
 601                               0, 0, _flags, _b, _ret)
 602
 603 static inline struct bkey_s_c bch2_btree_iter_peek_prev_type(struct btree_iter *iter,
 604                                                              unsigned flags)
 605 {
 606         return  flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 607                                                 bch2_btree_iter_peek_prev(iter);
 608 }
 609
 610 static inline struct bkey_s_c bch2_btree_iter_peek_type(struct btree_iter *iter,
 611                                                         unsigned flags)
 612 {
 613         return  flags & BTREE_ITER_SLOTS      ? bch2_btree_iter_peek_slot(iter) :
 614                                                 bch2_btree_iter_peek(iter);
 615 }
 616
 617 static inline struct bkey_s_c bch2_btree_iter_peek_upto_type(struct btree_iter *iter,
 618                                                              struct bpos end,
 619                                                              unsigned flags)
 620 {
 621         if (!(flags & BTREE_ITER_SLOTS))
 622                 return bch2_btree_iter_peek_upto(iter, end);
 623
 624         if (bkey_gt(iter->pos, end))
 625                 return bkey_s_c_null;
 626
 627         return bch2_btree_iter_peek_slot(iter);
 628 }
 629
 630 int __bch2_btree_trans_too_many_iters(struct btree_trans *);
 631
 632 static inline int btree_trans_too_many_iters(struct btree_trans *trans)
 633 {
 634         if (hweight64(trans->paths_allocated) > BTREE_ITER_MAX - 8)
 635                 return __bch2_btree_trans_too_many_iters(trans);
 636
 637         return 0;
 638 }
 639
 640 struct bkey_s_c bch2_btree_iter_peek_and_restart_outlined(struct btree_iter *);
 641
 642 static inline struct bkey_s_c
 643 __bch2_btree_iter_peek_and_restart(struct btree_trans *trans,
 644                                    struct btree_iter *iter, unsigned flags)
 645 {
 646         struct bkey_s_c k;
 647
 648         while (btree_trans_too_many_iters(trans) ||
 649                (k = bch2_btree_iter_peek_type(iter, flags),
 650                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 651                 bch2_trans_begin(trans);
 652
 653         return k;
 654 }
 655
 656 static inline struct bkey_s_c
 657 __bch2_btree_iter_peek_upto_and_restart(struct btree_trans *trans,
 658                                         struct btree_iter *iter,
 659                                         struct bpos end,
 660                                         unsigned flags)
 661 {
 662         struct bkey_s_c k;
 663
 664         while (btree_trans_too_many_iters(trans) ||
 665                (k = bch2_btree_iter_peek_upto_type(iter, end, flags),
 666                 bch2_err_matches(bkey_err(k), BCH_ERR_transaction_restart)))
 667                 bch2_trans_begin(trans);
 668
 669         return k;
 670 }
 671
 672 #define lockrestart_do(_trans, _do)                                     \
 673 ({                                                                      \
 674         u32 _restart_count;                                             \
 675         int _ret2;                                                      \
 676                                                                         \
 677         do {                                                            \
 678                 _restart_count = bch2_trans_begin(_trans);              \
 679                 _ret2 = (_do);                                          \
 680         } while (bch2_err_matches(_ret2, BCH_ERR_transaction_restart)); \
 681                                                                         \
 682         if (!_ret2)                                                     \
 683                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 684                                                                         \
 685         _ret2;                                                          \
 686 })
 687
 688 /*
 689  * nested_lockrestart_do(), nested_commit_do():
 690  *
 691  * These are like lockrestart_do() and commit_do(), with two differences:
 692  *
 693  *  - We don't call bch2_trans_begin() unless we had a transaction restart
 694  *  - We return -BCH_ERR_transaction_restart_nested if we succeeded after a
 695  *  transaction restart
 696  */
 697 #define nested_lockrestart_do(_trans, _do)                              \
 698 ({                                                                      \
 699         u32 _restart_count, _orig_restart_count;                        \
 700         int _ret2;                                                      \
 701                                                                         \
 702         _restart_count = _orig_restart_count = (_trans)->restart_count; \
 703                                                                         \
 704         while (bch2_err_matches(_ret2 = (_do), BCH_ERR_transaction_restart))\
 705                 _restart_count = bch2_trans_begin(_trans);              \
 706                                                                         \
 707         if (!_ret2)                                                     \
 708                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 709                                                                         \
 710         _ret2 ?: trans_was_restarted(_trans, _restart_count);           \
 711 })
 712
 713 #define for_each_btree_key2(_trans, _iter, _btree_id,                   \
 714                             _start, _flags, _k, _do)                    \
 715 ({                                                                      \
 716         int _ret3 = 0;                                                  \
 717                                                                         \
 718         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 719                              (_start), (_flags));                       \
 720                                                                         \
 721         while (1) {                                                     \
 722                 u32 _restart_count = bch2_trans_begin(_trans);          \
 723                                                                         \
 724                 _ret3 = 0;                                              \
 725                 (_k) = bch2_btree_iter_peek_type(&(_iter), (_flags));   \
 726                 if (!(_k).k)                                            \
 727                         break;                                          \
 728                                                                         \
 729                 _ret3 = bkey_err(_k) ?: (_do);                          \
 730                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 731                         continue;                                       \
 732                 if (_ret3)                                              \
 733                         break;                                          \
 734                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 735                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 736                         break;                                          \
 737         }                                                               \
 738                                                                         \
 739         bch2_trans_iter_exit((_trans), &(_iter));                       \
 740         _ret3;                                                          \
 741 })
 742
 743 #define for_each_btree_key2_upto(_trans, _iter, _btree_id,              \
 744                             _start, _end, _flags, _k, _do)              \
 745 ({                                                                      \
 746         int _ret3 = 0;                                                  \
 747                                                                         \
 748         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 749                              (_start), (_flags));                       \
 750                                                                         \
 751         while (1) {                                                     \
 752                 u32 _restart_count = bch2_trans_begin(_trans);          \
 753                                                                         \
 754                 _ret3 = 0;                                              \
 755                 (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, (_flags));\
 756                 if (!(_k).k)                                            \
 757                         break;                                          \
 758                                                                         \
 759                 _ret3 = bkey_err(_k) ?: (_do);                          \
 760                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 761                         continue;                                       \
 762                 if (_ret3)                                              \
 763                         break;                                          \
 764                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 765                 if (!bch2_btree_iter_advance(&(_iter)))                 \
 766                         break;                                          \
 767         }                                                               \
 768                                                                         \
 769         bch2_trans_iter_exit((_trans), &(_iter));                       \
 770         _ret3;                                                          \
 771 })
 772
 773 #define for_each_btree_key_reverse(_trans, _iter, _btree_id,            \
 774                                    _start, _flags, _k, _do)             \
 775 ({                                                                      \
 776         int _ret3 = 0;                                                  \
 777                                                                         \
 778         bch2_trans_iter_init((_trans), &(_iter), (_btree_id),           \
 779                              (_start), (_flags));                       \
 780                                                                         \
 781         while (1) {                                                     \
 782                 u32 _restart_count = bch2_trans_begin(_trans);          \
 783                 (_k) = bch2_btree_iter_peek_prev_type(&(_iter), (_flags));\
 784                 if (!(_k).k) {                                          \
 785                         _ret3 = 0;                                      \
 786                         break;                                          \
 787                 }                                                       \
 788                                                                         \
 789                 _ret3 = bkey_err(_k) ?: (_do);                          \
 790                 if (bch2_err_matches(_ret3, BCH_ERR_transaction_restart))\
 791                         continue;                                       \
 792                 if (_ret3)                                              \
 793                         break;                                          \
 794                 bch2_trans_verify_not_restarted(_trans, _restart_count);\
 795                 if (!bch2_btree_iter_rewind(&(_iter)))                  \
 796                         break;                                          \
 797         }                                                               \
 798                                                                         \
 799         bch2_trans_iter_exit((_trans), &(_iter));                       \
 800         _ret3;                                                          \
 801 })
 802
 803 #define for_each_btree_key_commit(_trans, _iter, _btree_id,             \
 804                                   _start, _iter_flags, _k,              \
 805                                   _disk_res, _journal_seq, _commit_flags,\
 806                                   _do)                                  \
 807         for_each_btree_key2(_trans, _iter, _btree_id, _start, _iter_flags, _k,\
 808                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 809                                         (_journal_seq), (_commit_flags)))
 810
 811 #define for_each_btree_key_reverse_commit(_trans, _iter, _btree_id,     \
 812                                   _start, _iter_flags, _k,              \
 813                                   _disk_res, _journal_seq, _commit_flags,\
 814                                   _do)                                  \
 815         for_each_btree_key_reverse(_trans, _iter, _btree_id, _start, _iter_flags, _k,\
 816                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 817                                         (_journal_seq), (_commit_flags)))
 818
 819 #define for_each_btree_key_upto_commit(_trans, _iter, _btree_id,        \
 820                                   _start, _end, _iter_flags, _k,        \
 821                                   _disk_res, _journal_seq, _commit_flags,\
 822                                   _do)                                  \
 823         for_each_btree_key2_upto(_trans, _iter, _btree_id, _start, _end, _iter_flags, _k,\
 824                             (_do) ?: bch2_trans_commit(_trans, (_disk_res),\
 825                                         (_journal_seq), (_commit_flags)))
 826
 827 #define for_each_btree_key(_trans, _iter, _btree_id,                    \
 828                            _start, _flags, _k, _ret)                    \
 829         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 830                                   (_start), (_flags));                  \
 831              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 832              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 833              bch2_btree_iter_advance(&(_iter)))
 834
 835 #define for_each_btree_key_upto(_trans, _iter, _btree_id,               \
 836                                 _start, _end, _flags, _k, _ret)         \
 837         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 838                                   (_start), (_flags));                  \
 839              (_k) = __bch2_btree_iter_peek_upto_and_restart((_trans),   \
 840                                                 &(_iter), _end, _flags),\
 841              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 842              bch2_btree_iter_advance(&(_iter)))
 843
 844 #define for_each_btree_key_norestart(_trans, _iter, _btree_id,          \
 845                            _start, _flags, _k, _ret)                    \
 846         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 847                                   (_start), (_flags));                  \
 848              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 849              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 850              bch2_btree_iter_advance(&(_iter)))
 851
 852 #define for_each_btree_key_upto_norestart(_trans, _iter, _btree_id,     \
 853                            _start, _end, _flags, _k, _ret)              \
 854         for (bch2_trans_iter_init((_trans), &(_iter), (_btree_id),      \
 855                                   (_start), (_flags));                  \
 856              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),\
 857              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 858              bch2_btree_iter_advance(&(_iter)))
 859
 860 #define for_each_btree_key_continue(_trans, _iter, _flags, _k, _ret)    \
 861         for (;                                                          \
 862              (_k) = __bch2_btree_iter_peek_and_restart((_trans), &(_iter), _flags),\
 863              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 864              bch2_btree_iter_advance(&(_iter)))
 865
 866 #define for_each_btree_key_continue_norestart(_iter, _flags, _k, _ret)  \
 867         for (;                                                          \
 868              (_k) = bch2_btree_iter_peek_type(&(_iter), _flags),        \
 869              !((_ret) = bkey_err(_k)) && (_k).k;                        \
 870              bch2_btree_iter_advance(&(_iter)))
 871
 872 #define for_each_btree_key_upto_continue_norestart(_iter, _end, _flags, _k, _ret)\
 873         for (;                                                                  \
 874              (_k) = bch2_btree_iter_peek_upto_type(&(_iter), _end, _flags),     \
 875              !((_ret) = bkey_err(_k)) && (_k).k;                                \
 876              bch2_btree_iter_advance(&(_iter)))
 877
 878 #define drop_locks_do(_trans, _do)                                      \
 879 ({                                                                      \
 880         bch2_trans_unlock(_trans);                                      \
 881         _do ?: bch2_trans_relock(_trans);                               \
 882 })
 883
 884 #define allocate_dropping_locks_errcode(_trans, _do)                    \
 885 ({                                                                      \
 886         gfp_t _gfp = GFP_NOWAIT|__GFP_NOWARN;                           \
 887         int _ret = _do;                                                 \
 888                                                                         \
 889         if (bch2_err_matches(_ret, ENOMEM)) {                           \
 890                 _gfp = GFP_KERNEL;                                      \
 891                 _ret = drop_locks_do(trans, _do);                       \
 892         }                                                               \
 893         _ret;                                                           \
 894 })
 895
 896 #define allocate_dropping_locks(_trans, _ret, _do)                      \
 897 ({                                                                      \
 898         gfp_t _gfp = GFP_NOWAIT|__GFP_NOWARN;                           \
 899         typeof(_do) _p = _do;                                           \
 900                                                                         \
 901         _ret = 0;                                                       \
 902         if (unlikely(!_p)) {                                            \
 903                 _gfp = GFP_KERNEL;                                      \
 904                 _ret = drop_locks_do(trans, ((_p = _do), 0));           \
 905         }                                                               \
 906         _p;                                                             \
 907 })
 908
 909 /* new multiple iterator interface: */
 910
 911 void bch2_trans_updates_to_text(struct printbuf *, struct btree_trans *);
 912 void bch2_btree_path_to_text(struct printbuf *, struct btree_path *);
 913 void bch2_trans_paths_to_text(struct printbuf *, struct btree_trans *);
 914 void bch2_dump_trans_updates(struct btree_trans *);
 915 void bch2_dump_trans_paths_updates(struct btree_trans *);
 916
 917 struct btree_trans *__bch2_trans_get(struct bch_fs *, unsigned);
 918 void bch2_trans_put(struct btree_trans *);
 919
 920 extern const char *bch2_btree_transaction_fns[BCH_TRANSACTIONS_NR];
 921 unsigned bch2_trans_get_fn_idx(const char *);
 922
 923 #define bch2_trans_get(_c)                                              \
 924 ({                                                                      \
 925         static unsigned trans_fn_idx;                                   \
 926                                                                         \
 927         if (unlikely(!trans_fn_idx))                                    \
 928                 trans_fn_idx = bch2_trans_get_fn_idx(__func__);         \
 929         __bch2_trans_get(_c, trans_fn_idx);                             \
 930 })
 931
 932 void bch2_btree_trans_to_text(struct printbuf *, struct btree_trans *);
 933
 934 void bch2_fs_btree_iter_exit(struct bch_fs *);
 935 int bch2_fs_btree_iter_init(struct bch_fs *);
 936
 937 #endif /* _BCACHEFS_BTREE_ITER_H */